Stop Slopping your Skills - Rahul Baboota

Rahul Baboota

SubscribeSign in

Stop Slopping your Skills

Rahul Baboota<br>Jun 22, 2026

Share

Skills are one of the hottest things right now and rightfully so. They are a very elegant way to manage context windows as well as provide specific functionality to agents.<br>A big reason for skills getting massively popular is because of how straightforward (yet powerful) it is to build one. Since the simplest of skills can be expressed in natural language, even non-technical people can use it easily and effectively.<br>BUT ...<br>As with every other thing recently that gains a bit of traction, there has been a huge influx of skill slop where everyone is building and sharing 1000 line skills that are a total game changer.<br>To put fuel to this fire, almost every harness now supports a skill-creator skill where people can simply prompt their agent to build a skill to do X without ever having an actual look at the produced skill (this is BAD).<br>Now, I’m not against this skill-creator tooling as it’s a great way of enabling people to build skills. But what I am against is blindly using this function to generate more slop.<br>There are some things in which effort put into human curation can go a very long way and skills are that.<br>Here is a no-fluff guide to help you build better skills based on my experience:<br>Use Description Correctly

One of the most important things to know is that the description is the routing trigger and not a summary of what the skill does. This is a very important distinction.<br>It’s one of the most common points of failure I’ve seen. Most people treat the skill description as an internal documentation for what it does. A good description specifies when the agent should load the skill, not what it does.

Here’s a quick checklist for writing a good description:<br>Use phrases like “Load when”, “Trigger on”

Target 50 words or fewer

Describe the user’s intent, ideally from real queries

Use Progressive Disclosure

A skill is not just one massive file, it’s a folder which can have:<br>Scripts

References

Templates

If there is any portion of your skill that’s conditional or extremely heavy in content, take it out of the main file and put it in a different reference file.

Using scripts is another massive unlock. If there is any part of your skill where the agent is writing the same code every time, package it up in a script and give the main file a pointer to it. This makes your task much more deterministic and also saves up a bunch of tokens.<br>These files won’t clutter your context window because of progressive disclosure. Your agent will reference these files and bring them in only as it needs them. This is a much better design than stuffing everything into one giant file.<br>Have a Gotcha’s Section

Skills do not end at creation. This is something that I have learned the hard way. Just like software, skills need to be maintained as well.<br>Initially your skill is going to work ~80% of the time. But the thing that will take it to 99% is the Gotcha’s section. These are the weird failure cases your agent is likely to mishandle unless you explicitly warn it.<br>So, whenever you notice a specific failure case in testing/production, add it to this list. It’s a negative example so it’s not explicitly changing the skill behavior but it lets the model know: “Hey, there’s this known failure”.

As you keep using your skill, your gotchas list should ideally keep growing from real failures making them one of the highest-value parts of your skill.<br>In conclusion, It’s hard to write a short skill. If your skill is easy to write, it is probably too long or shouldn’t exist. A good skill is as short as it can be.<br>The act of building skills makes you better at building more skills.

Share

Discussion about this post<br>CommentsRestacks

TopLatest

No posts

Ready for more?

Subscribe

© 2026 Rahul Baboota · Privacy ∙ Terms ∙ Collection notice<br>Start your SubstackGet the app<br>Substack is the home for great culture

This site requires JavaScript to run correctly. Please turn on JavaScript or unblock scripts